[mpiwg-persistence] Persistent collectives progress update

[HOLMES Daniel]

FYI

Cheers,
Dan.

Begin forwarded message:

From: Dan Holmes <dholmes at epcc.ed.ac.uk<mailto:dholmes at epcc.ed.ac.uk>>
Date: 3 October 2017 at 13:50:10 BST
To: "Sridharan, Srinivas" <srinivas.sridharan at intel.com<mailto:srinivas.sridharan at intel.com>>

Hi Srinivas,

The paper presentation went very well. There is definitely community interest and support for this effort. Two of the other papers presented detailed implementation optimisations that require the persistent collective interface, one from a theoretical point-of-view with a proof-of-concept code, and the other with a more practical approach using the profiling intercept mechanism to override the existing MPI implementation of particular collective functions. Both demonstrated improved performance derived from performance planning - perfect use-cases for our interface work.

The formal reading at the MPI Forum prompted a variety of small wording changes to the proposed text and a couple of small semantic changes - at least one which requires further design work in the working group. This will likely reset the procedure, i.e. we will probably be required to formally read the modified proposal again at the next meeting with the additional semantic change included. That issue is around freeing the persistent collective request object - the current MPI_REQUEST_FREE function is local (does not rely on MPI calls elsewhere, specifically this means *not* collective) and nonblocking (it may not have actually freed resources when it returns to user code). There is a good argument for adding a new function that gives a stronger guarantee and requires a stronger usage model, i.e. a blocking collective request free function. There is a design space of 8 or 12 possible combinations of the fundamental semantics, which we should explore.

Tony/Bradley: I will let you guys comment on where we are at with sharing our reference implementation code with Srinivas and his team.

Cheers,
Dan.

On 3 Oct 2017, at 13:30, Sridharan, Srinivas <srinivas.sridharan at intel.com<mailto:srinivas.sridharan at intel.com>> wrote:

Hi Dan,

Any interesting comments on the paper at EuroMPI?

Also, seems like there was supposed to be a reading of persistent collectives. How did it go?
http://mpi-forum.org/meetings/2017/09/agenda
Finally, please let me know if you have any updates on sharing your code with me…

Srinivas
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-persistence/attachments/20171004/d7fcd2ce/attachment-0001.html>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: not available
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-persistence/attachments/20171004/d7fcd2ce/attachment-0001.ksh>